UNSUPERVISED FEATURE LEARNING FOR MID-LEVEL DATA REPRESENTATION

Authors : Emrah Ergül, Mehmet Karayel, Oğuzhan Timuş, Erkan Kiyak, Turkish Naval Academy

Pages : 51-79

View : 17 | Download : 8

Publication Date : 2016-04-24

Article Type : Other

Abstract :Attribute based approaches are commonly used in recent years instead of  low level features for image classification which is one of the most important problems in the field of computer vision. The most important advantage of attribute based approach is that learning can be performed similar to human by using attributes which makes sense for people. In this study, unsupervised attributes are developed in order to avoid human related problems in supervised attribute learning. In our proposed work, the attributes are generated as random binary and relative definitions. The process of random attribute generation simplifies the data modeling when compared to other work in the literature. In addition, a major problem which is the increasing the numbers of attributes in attribute based approaches is eliminated owing to the increasing the numbers of attributes easily. Furthermore, attributes are selected more wisely using simple applicable algorithm to improve the discriminative capacity of randomly generated attribute set for image classification. The proposed approaches are evaluated with the other similar attribute based studies comparatively in the literature based on the same data set (OSR-Open Scene Recognition). Experiments show that noteworthy performance increase is achieved.
Keywords : Relative Attributes, Unsupervised Feature Extraction, Attribute Selection, Visual Recognition